The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
Increasing Recommendation Accuracy and
Diversity via Social Networks Hyperbolic
Embedding
Vasiliki Pouli
University of Maryland
Institute for Systems Research,
College Park, MD 20742, USA
vpouli@umd.edu
John S. Baras
University of Maryland
Institute for Systems Research,
College Park, MD 20742, USA
baras@umd.edu
Abstract—Several applications are built around sharing
information by leveraging social network connections. For
example, in social buying sites like Groupon, a deal is
usually forwarded to interested recipients through their
social graph. A primary goal is to improve user satisfaction
by maximizing the relevance of the shared message to the
target audience. In order to suggest more personalized
products, one should consider offering not only accurate
but also diverse recommendations, since diversification
plays an important factor in increasing the users's satisfaction. In this work, we address this problem by proposing a
social network hyperbolic embedding that exploits both social connections and user preferences aiming at increasing
both the accuracy and the diversity of recommendations.
I. INTRODUCTION
The increase of online commercial environments offers users millions of choices, which makes it hard
and time consuming to evaluate them in order to select
a service or product. Moreover, promoting products is
often considered as spam, especially if the recommended
products are irrelevant to the recipient. Thus, a primary
challenge is how to effectively recommend a product
only to users that are the most likely to find this
product interesting, with the goal to increase both user
satisfaction and profit [20].
For this purpose, several modern applications could
leverage social network connections in order to promote products. For example, social buying sites like
Groupon.com or auction websites such as eBay.com
could use a very effective way to promote their products
or services by sharing product recommendations to users
and their friends via social networks such as Facebook
[8], Twitter [3] etc.
Recommender systems are used to support the user
decision-making process by providing suggestions to
users about what to buy, which movie to watch, what
978-1-4799-2355-7/14/$31.00 ©2014IEEE
Anastasios Arvanitis
National Technical University of Athens,
Iroon Polytechniou, Zografou 157 80
Athens, Greece
anarv@dblab.ntua.gr
to listen, where to travel, etc. For this reason they have
become a valuable tool for online platforms to help customers in their planning and purchasing decisions. For
example in eBay or Amazon the platform recommends
the most relevant items to each user based on her search
queries and the behavior of like-minded users. Another
example is movie rental applications such as Netflix,
where the platform suggests movies according to users'
previous ratings and other users' similar tastes.
Which factors have to be considered when recommending things? And which are the most successful
techniques to generate recommendations to interested
users, such that we avoid spamming everyone with
irrelevant information? The problem of providing accurate recommendations to users is generally complex;
Netflix announced the Netflix Prize competition (netflixprize.com) [4] with a prize of $1M for those who would
improve the accuracy of predictions by 10%.
Recommendations can be done via various methods,
however among the most well-known methods is Collaborative Filtering (CF) [6]. CF is a method that can
be used to select items which are similar to other items
that a user has rated (item-based CF), as it is used in
Amazon.com [14], [16] or by utilizing information collected from users with similar interests (user-based CF)
[21], as seen in GroupLens. The first step of collaborative
filtering measures the similarity of users/items based on
various similarity measures such as cosine similarity or
pearson correlation that we will discuss in Section 2. The
k users/items that have the highest similarity with the
active user/item are selected; these are usually referred
to as neighbors. Finally, the ratings of the neighbors are
aggregated and used to predict the user's rating.
In order to improve the accuracy of recommendations,
one has to consider how relevant a product or service
225
The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
is to a user. This relevance score is often contextsensitive; i.e., it depends on several factors represented
as dimensions. Such dimensions could be for example
the thematic category, price, location of the product. If
we consider, e.g. a coupon for a Chinese restaurant in
Washington DC, it would be wise to forward it mainly
to users that are located near Washington DC and are
interested in Chinese cuisine. Thereby, it is more likely
that these users will accept the recommended deal and
eventually buy it. On the other hand, flooding a message
concering a product or service to social connections of a
user, including those that live far away from Washington
DC or have not expressed any interest for food coupons,
would probably be considered as spam. The dependence
of relevance to the message context [18] calls upon
developing a dynamic routing approach.
Moreover, when recommending products or items an
important factor that should be also taken into account
apart from accuracy is diversity [15], [22]. The importance in having diverse recommendations has been highlighted in several studies (e.g. [1], [2], [5]). Increasing
the diversity in recommendations can help companies
sell a variety of products in addition to the highly ranked
ones. It could also be beneficial for some providers that
want to promote long tail items, i.e. items that are not
so popular in the sales distribution either because they
are older or because they are low budget ones.
In this work we propose building a hyperbolic embedding of a user friendship network by bringing similar
users closer together. This will ensure message delivery
to all the recipient nodes of the social network and
produce better accuracy and diversity compared to the
traditional CF techniques. The model considers user
preferences and delivers product recommendations to the
nodes that are the most interested, thus offering better
personalization to the users.
II. BACKGROUND INFORMATION AND RELATED
WORK
A. Social-based Routing Algorithms
Recent works [10], [11], [12] have proposed routing
algorithms that use social characteristics with applications on delay tolerant, ad hoc or P2P networks. In
[12] the authors propose a socially-based greedy routing
algorithm for delay tolerant networks. They use an ndimensional social profile of users and utilize the Jaccard
coefficient to measure the similarity between them. Then,
they apply a greedy routing algorithm by selecting the
nodes that are socially closer to each other. Also, in [11]
the authors proposed the Bubble rap, where they utilized nodes' centralities and communities to improve the
forwarding efficiency. Furthermore, in [10], the authors
have shown that labeling the nodes with their affiliation
and forwarding messages to nodes belonging to the same
affiliation as the destination, can improve forwarding
performance both in terms of delivery ratio and cost.
However, although these algorithms exploit user-to-user
similarities, they (i) do not utilize their preferences in
terms of the forwarded message content, they (ii) do
not ensure message delivery to all interested recipients
and they (Hi) do not take into account social friendship
connections. Motivated by this, in our current work we
propose a context-based routing approach that exploits
both user preferences along with the social network
friendship connections and creates a network structure
aiming at maximizing relevance and potential profit by
ensuring greater accuracy and diversity in recommendations to the most interested nodes.
B. Recommendation Systems.
Recommendation systems have proved very helpful
for users that wish to find interesting items from a very
large information and product space. One of the most
common methods used is collaborative filtering (CF)
where similar users or items are identified and then the
top-k most highly ranked among them are selected to be
used for final predictions. The basic idea for collaborative filtering lies in calculating the similarity between
each user/item by comparing the common ratings of
each user/item with other users/items. The most common
similarity measures used in collaborative filtering are:
• Cosine similarity where user/item ratings are represented as points in a vector space model and cosO
is calculated between them. Depending on whether the
similarity is measured on items or users, the cosine
similarity can be either item-based or user-based. Itembased cosine similarity is computed by considering corated items only. These co-rated pairs are obtained from
different users. Item-based cosine similarity is defined as
follows:
E
■ ,. -n
simii.i
) = —,
R(u, i) * R(u, i')
ueU(i,i')
—,
E
R(u,i')>
Y ueU(i,i')
(1)
where sim{i,i') is the similarity of item i with item
i',R(u,i) is the rating of user u for item i,R(u,i') is
the rating of user u for item i' and U(i,i') is the set of
all users that have rated both items i and i' as seen in
Figure 1.
User-based cosine similarity can be computed in a
similar way by interchanging items with users.
226
/ E
RM*
Y ueU(i,i')
The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
•Pearson correlation coefficient: The similarity between items can be also computed based on the Pearson correlation which measures the linear relationship
between objects and is calculated using the following
equation:
sim(i,i') =
£
ueU(i,i')
2
R(l,1>
^h
-
W /
k
u
1
2
(R(u,i)-R(i))*(R(u,i')-R(i'))
uEU(i,i')
£
1
(fl(«,0- ■ my
ueU(i,i')
1
R{i'))2
(2)
where sim(i,i') is the similarity of item i with item
i', R(u, i) is the rating of user u for item i, R(u, i') is
the rating of user u for item i', R(i) is the mean rating of
item i, R{i') is the mean rating of item i' and U(i, i') is
the set of all users that have rated both items i and i' as
seen in Figure 1. The Pearson Correlation measures the
strength of the linear relationship between two variables
and takes values in the range [-1, 1], where -1 shows a
negative correlation, 0 means no correlation, and 1 is a
positive correlation.
Several variations of the above similarity measures
have been also proposed in the literature including
adjusted weighted schemes, Spearman rank correlation,
Kendall's Correlation, mean squared differences, entropy [19]. Spertus et al. [17] conducted a large-scale
study to evaluate six different similarity measures on the
Orkut social network and their experiments showed that
the best results in recommendations were those produced
based on the cosine similarity.
Although accuracy is usually the target of the recommendation systems [4], the produced recommendations
are not always useful for the users who often wish for
something different or diverse. Thus, diversity should
also be taken into account [1], [5], [15]. With this
goal, Bradley and Smyth [5] have proposed a method to
improve recommendation diversity by considering it as
the average dissimilarity between all pairs of items in the
result-set. We will use the diversity metric as proposed in
[1], [2]. In these works Adomavicius and Kwon adopt
CF ranking-based heuristics to improve the aggregate
recommendation diversity without affecting the accuracy.
They achieve this by recommending items whose ratings
are above a specific threshold. In our work, we consider
the same diversity measure but examine a different
accuracy metric that does not take into account the
rating thresholds, since it sometimes limits the available
options. We create a hyperbolic network and analyse
its recommendation performance by comparing it with
traditional CF techniques. We show that our model produces better accuracy and diversity measures compared
to the traditional CF methods, when forwarding relevant
j
*K
R(i.k)
\R(ly
m-1
m
Fig. 1. User (columns) - Item (rows) rating matrix
product recommendations to the most interested social
connections.
III. RECOMMENDATIONS IN HYPERBOLIC ALLY
EMBEDDED SOCIAL NETWORKS
In this section we will describe our similarity measure
that is used for calculating node-message relevance distances. Our similarity measure consists of two parts; the
first part incorporates the context similarity, whereas the
second part measures the (structural) network distance.
Subsequently, we will present our context-based routing
algorithm that utilizes our proposed node similarity measure in order to propagate the message to all interested
recipients.
A. User-item context model
We will first explain the context part by describing the
context model that we used. We will assume that both the
product message and a users' interests/preferences can be
represented using the same context model. This context
information can be modeled as a vector of attributes
where each attribute represents a context dimension. In
general, the domain for each context dimension might
take values from a hierarchy, such as the one depicted in
Figure 2. Such context dimensions could be for example
the content topic (e.g. food, entertainment, traveling,
movie genre), a location, price range, etc.
We assume that each user has specific interests which
can be defined using a set of keywords. These keywords
227
The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
Fig. 3. User similarity network
Fig. 2. A Context Hierarchy
might take any value from the attribute domain. For
example a user, as seen in Figure 2, might have specified
that she likes 'Thai' cuisine and 'Action' movies in
'Baltimore, MD' . These preferences can be explicitly
provided by each user (such as Facebook 'likes') or
they might as well have been extracted by applying data
mining techniques on the user's posts, comments, shares
etc. We also consider a product item that has values
taken from the same context model with a product vector
Ij= {Asian, Baltimore}, which indicates that it refers
to a deal about Asian food in a restaurant located in
Baltimore, MD. We could also define weights in the
context dimensions and say for example that a user
likes Asian cuisine with a weight score of 80%. As
user preferences we could use the items (e.g. books,
movies, music) that users recommend to their friends
in a social network such as the Douban.com Chinese
network. In such a network a user can become a friend
or fan of another user, collect interesting movies, books
and music albums, rate them or recommend them to his
friends and fans. The friends of a user can see a social
recommendation and either accept it by collecting the
item later, or choose to decline it.
In order to measure the context similarity between a
message and a user's preferences, we will use a cosine
similarity between the preference vector of user I/, and
the product vector Ij.
context_similarity(Ui,
Ij) = ,, ' . 2 ..
ll"»llllfjll
(3)
network. We will try to identify who will be the most
similar among them according to their preferences. We
call our new similarity measure as follower's similarity
measure and we define two versions: one that takes into
account the user ratings (rating-based) and the other
one that does not take ratings into account (non ratingbased).
1) Non Rating-based: : Usually in a social network
nodes are connected with friendship connections or with
connections that specify who follows whom where nodes
follow each other based on their common interests
or ratings. We adapt these 'follower' connections by
introducing some weights according to the following
rule: two connected users have k edge weight if they
have rated k common items, k will be the edge weight
and similarity measure between them, normalized when
divided by the total number of items thus:
sim(u, u') =
In this section we introduce a new similarity measure
to estimate the similarity between nodes in a social
'
= k/mt
(4)
For example, in Figure 3 where we assume a movie
recommendation system where users can rate the movies
they have watched, users 1 and 2 are connected with
weight 1 since both of them have rated one movie (Ml).
2) Rating-Based- : In the second version we take
into account the user ratings for the items and compute
the edge connection weight between the two users as
disproportional to the difference of the ratings for the
common items they have rated. The more the users differ
in their ratings the more dissimilar they are.
sim(u, u') =
B. User-user follower Similarity Measure
|J(M M/)I
—;—
(5)
+ z2iei(u,u') \R(U> *) " R(u > *)I
where I(u,u') is the set of items rated by both user u
and v!.
228
1
The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
For example user 1 is connected with user 4 since
both of them have rated movie 5 (M5) (1 common
movie). Note that user 1 has also rated movie 1 (Ml) and
movie 2 (M2) and user 4 has rated movie 4 (M4). These
movies might not be similar; however the fact that both
users 1 and 4 have rated M5 suggests that they might
be interested in each other's different choices (M2 and
M4) as well and follow them in order to increase the
diversification of their recommendations.
the items that were rated as top-k from the original user
ratings, across all users:
C. Recommendation Predictions
By creating user-user follower networks and the edge
weights between each two nodes, we infer their similarity and use it as a similarity measure instead of
Equation 1 to predict ratings as described below in
Equation 6.Using our model a network with weighted
edges is created and users can follow their most similar
neighbors' top recommendations to increase the diversity
of their recommendations. In order to estimate the rating
R*(u,i) that user u would assign to an item i we
compute R*(u, i), which is the adjusted weighted sum
of all ratings R(u', i), where u' belongs to K[u): the K
neighbors of user u who rated item i and have the highest
similarity sim(u,u') to user u. R(u) is the average
rating of user u and R(u') of user u' respectively. This
weighted sum can capture how the users rate common
items and is defined as follows [2]:
E. Hyperbolic Embedding
£
sim(u,u')(R(u',i) - R(u'))
u'EK{u)
(6)
D. Accuracy vs. Diversity
As discussed above, by increasing the diversity in recommendations we could have more personalized items
and companies could also increase their revenues by offering long tail products. Users could also have diversity
in taste by following their friends actions. We would like
to measure the accuracy and diversity variations under
different scenarios.
By accuracy, as in [2], we measure the statistical precision measure, which we consider as the percentage of
items from the top-k high-ranked predictions measured
from (6) which are indeed rated as the top-k in the
original ratings of each user.
accuracy@k = \correct(Sk(u))\/ k
(7)
where Sk(u) is the set of the k highest ranked predicted
items for user u.
By diversity we define the union of the different top-k
highest rated items predicted from eq. (6) compared to
diversity@k = \ Sk (u) - S'k (u) |
(8)
where Sk(u) is the set of the k highest ranked predicted
items for user u and S'k{u) is the set of the k items that
were originally ranked as top — k in the original user
ratings of user u.
Furthermore, in order to ensure message delivery to
all interested recipients, our approach greedily embeds
the network into the hyperbolic space. We followed
the greedy hyperbolic embedding since, as it is shown
in [13], every finite, connected, undirected graph has
a greedy embedding in a two-dimensional hyperbolic
space, i.e., one may achieve 100% success rate with
greedy routing by assigning virtual coordinates in
the hyperbolic plane rather than the Euclidean plane.
Based on this embedding, each node will always have
a neighbor closer to the destination and a message will
never get stuck into local minima as in Euclidean space.
In order to embed the network into the hyperbolic
space we applied the algorithm described in [7].
Following this approach, we constructed a spanning
tree from the original graph, and assigned to each node
hyperbolic coordinates from the set B ={z G C||z|< 1}
in accordance to their parents coordinates. A greedy
embedding of the spanning tree is also a greedy
embedding of the graph. We consider as network
distance network_dist(z\,Zi) between two nodes zi and
Z2 the hyperbolic distance between them. Figure 4
shows a random graph of 20 nodes and Figure 5 its
minimum spanning tree hyperbolic embedding in the
Poincare half plane disc.
F. Routing Algorithm
The goal of the routing algorithm is to ensure that
each message will reach, in the minimum number of
steps, the nodes that are most relevant to a message.
We will assume that a product vendor wants to forward
the message through his friends/followers' connections
to the top-m most interested neighbors in the minimum
number of steps. Based on the proposed context similarity measure and hyperbolic embedding mapping, we
defined a relevance metric, named: relevance, given by
Equation 9. This metric incorporates both the context
similarity and the (structural) network distance between
user i and u' and calculates the relevance of a node i to
an item j :
229
The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
nodes i.e. the users where the message will be forwarded
to, are the ones that maximize our similarity metric:
relevance(i,j).
After identifying the target nodes, we proceed to our
routing algorithm. Initially, the algorithm begins from
D, i.e., the node that issues the message. For example
in Facebook [9], D might be the account of a service
provider. Next, at each step the algorithm utilizes our
network embedding to find out which of the neighbors
are closer to the user than the others. From these nodes,
the algorithm selects to forward the recommendation to
the n neighbors that have the highest relevance score
based on Equation 9.
)
l
j
IV. EXPERIMENTAL EVALUATION
J
y
We conducted a set of experiments to examine the
efficiency of our algorithm which chooses nodes most
relevant to the message against a non social-based
scheme, where nodes choose neighbors to forward the
message to without taking into account their preference
similarity. We used both synthetic and real data to test
the efficiency of our model.
,
A. Synthetic Network
Fig. 4. Random graph with 20 nodes
)anningtree
Fig. 5. Greedy Hyperbolic embedding of the minimum spanning tree
re disc
of the graph depicted in Figure 4 in the half plane Poincare
disc
argmax relevancefi, j)
i
T ,.
context_similarity(Ui, —i^Ij) (9)
network_dist(i, u')
In order to discover these nodes, we cann simply
work and
evaluate Equation 9 on the nodes of the network
loser and
identify the nodes that are hyperbolically closer
most similar to the source node with regard to the useritem context similarity measure along with the user-user
similarity measure and the network distance available
he target
through the hyperbolic graph embedding. The
For the synthetic network we assume that a product
provider who wants to recommend a product could
specify under which categories and sub-categories the
product belongs to, e.g., Asian food, thriller movies, traveling to Hawaii etc. The subscribers could as well have a
similar menu that would allow them to select for which
categories they would like to receive messages/products.
Thereby, the system will receive each item and forward it
to the recipients that have shown interest to the particular
categories associated with the product.
^or ^ c r e a t ion of the synthetic network we used
the R Project (www.r-project.org) and generated an
undirected random network with AT=100 nodes and
probability for adding an edge between two arbitrary
vertices equal to 0.02. We chose m=l meaning that
we take one destination i.e. the node i that maximizes
relevance(i,j). Further, we also used a fanout = 1 i.e.,
our routing algorithm will pass a recommendation to the
one neighbor that maximizes his context similarity with
the product and has a small network distance from the
destination node.
In each experiment we selected random user preference and product values for the context dimension
categories and we executed the simulation 100 times by
selecting a different source node for every iteration,
Through our simulations we want to compute the path
preference pathPref which we define as the average
230
The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
0.9
cf=30
U.8
0.4
0.3
155
160
165
170
175
180
185
190
195
Diversity
«^cosine
^^^movienetR
movienetNoR
Fig. 7. Diversity vs Accuracy
# of simulations
Fig. 6. Path preference
context similarity of the coupon with all the nodes along
the path:
M
context_sim(i, j)
(10)
pathPref = VJ
M
where M is the number of hops along the path.
The motivation for using path preference as a metric
in our experiments, is that we want to evaluate the
relevance of the message with the users that belong to
the path followed. The closer a user is to a producer
node in the social graph, the more likely it is to like
the products he offers, which increases the probability
of buying those products. Thus, path preference metric
could be considered as a profit measure.
We calculate the path preference from the source to
the destination using our proposed context-based routing
algorithm, and then we compare it against a greedy
forwarding routing which selects to forward a message
recommendation to the neighbor closest to the destination but without taking into account the preferences of
the nodes.
According to Figure 6 we observe that by exploiting
the user preferences in our algorithm, then the whole
path preference increases by 94%, compared to the
simple greedy case, which does not take into account
the similarity of the user with the product.
B. Evaluation of Recommendation Diversity and Accu­
racy in Real Network Data
To test the effectiveness of our model in a real
network we executed experiments using a real dataset
provided by the Douban.com social site. The dataset
consists of 2000 items rated by 1000 users. From eq. (6)
we compute the top-3 highly predicted items for each
user. The algorithm then tests how the mean accuracy
and the overall diversity varies for the top-3 highly
predicted items forwarded to the most relevant node, that
is the most relevant node for a specific source user and
forwarded item. We measure the average accuracy along
all users in the path as the mean of the accuracies defined
in eq. (7) across all users and the overall diversity as the
union of the diversities defined in (8) for all users. The
algorithm starts from random source nodes each time
and follows different routes depending on the relevance
of each item to the users. To find the most relevant node
for each item we used three different similarity metrics,
namely the user-based cosine similarity - eq. (1) (seen as
'cosine' in Figure 7) and the other two by considering the
proposed non rating-based ('followerNoR') and ratingbased ('followerR') user-follower similarity network eq. (4) and (5). In each scenario, as seen in Figure 7,
we used an increasing number of neighbors ranging
from 30 to 90 with a step of 30 each time. From the
measurements we can see the tradeoff between accuracy
and diversity in these three scenarios: the accuracy
is high when diversity is low and then decreases as
the diversity increases; meaning that the more accurate
recommendations we have the less diverse options we
offer to the users. This is normal since the more accurate
recommendations are typically safer and closer to the
user's expressed interests.
Further, we can observe that the two versions of our
proposed user-follower network exhibit better diversity
than the conventional cosine similarity. We chose to test
our model for various numbers of neighbors to show how
it affects the tradeoff between the accuracy and diversity.
We found out that the more neighbors selected the less
accuracy is achieved but with greater diversity, since the
sources of recommendations are wider by selecting a
greater amount of people that provide options not so
231
The 11th Annual IEEE CCNC - Multimedia Networking, Services and Applications
similar to a user. From the measurements we can see
that compared to the conventional cosine similarity, the
two versions of the user follower network have better
diversity and accuracy for the same number of neighbors
used.
[2] G. Adomavicius and Y. Kwon. Improving aggregate recommendation diversity using ranking-based techniques. IEEE TKDE,
24(5):896-911, 2012.
[3] I. Anger and C. Kittl. Measuring influence on twitter. In I-KNOW,
page 31, 2011.
[4] J. Bennett and S. Lanning. The netflix prize. In In KDD Cup
and Workshop in conjunction with KDD, 2007.
[5] K. Bradley and B. Smyth. Improving recommendation diversity.
In Proc, 12th Irish Conf. Artificial Intelligence and Cognitive
Science, 2001.
V. CONCLUSION
In this paper we proposed a context-aware routing
scheme that aims to increase the relevance of messages
shared across a social network. This is achieved by forwarding the message to the most relevant nodes, taking
into account both user preferences and similarity along
with the network structure. Our experiments show that
our context-aware routing scheme provides significant
improvements in terms of the preference similarity of
nodes along the path that is followed, thus increasing
the chances of users that might be interested to the
message context to buy a product. In order to identify the
similar users we described a social connection network
hyperbolic embedding that aims at creating weighted
connections between users to increase the overall diversity and accuracy of the items recommended to most
interested and similar nodes. Our model is flexible since
it can be adopted by the widely used CF technique and
provides a follower's connection measure for selecting
the right neighbors according to the system needs. If
we aim at improving the accuracy we should create
a rating-based connection between users. On the other
hand, if greater diversity is required then the non-rating
based approach should be followed. We adopted the
steps of the CF technique to make predictions and saw
that, compared to the traditional cosine similarity used
to select neighbors, our proposed approach yields better
results by offering greater accuracy and diversity.
ACKNOWLEDGEMENT
Research partially supported by the US Air Force
Office of Scientific Research MURI grants FA9550-101-0573 and FA9550-09-1-0538.
REFERENCES
[1] G. Adomavicius and Y. Kwon. Overcoming Accuracy-Diversity
Tradeoff in Recommender Systems: A Variance-Based Approach.
In Proc of the 18th Workshop on Information Technology and
Systems (WITS'08), 2008.
[6] J. S. Breese, D. Heckerman, and C. M. Kadie. Empirical analysis
of predictive algorithms for collaborative filtering. In UAI, pages
43-52, 1998.
[7] A. Cvetkovski and M. Crovella. Hyperbolic embedding and
routing for dynamic graphs. In INFOCOM, pages 1647-1655,
2009.
[8] N. B. Ellison, C. Steinfield, and C. Lampe. The benefits of
Facebook "friends: " social capital and college students' use of
online social network sites. J. Computer-Mediated Communica­
tion, 12(4): 1143-1168, 2007.
[9] J. Feiler. How to do everything : Facebook applications.
McGraw-Hill, New York, 2008.
[10] P. Hui and J. Crowcroft. How small labels create big improvements. In PerCom Workshops, pages 65-70, 2007.
[11] P. Hui, J. Crowcroft, and E. Yoneki. Bubble rap: Social-based
forwarding in delay-tolerant networks. IEEE Transactions in
Mobile Computing, 10(11): 1576-1589, 2011.
[12] K. Jahanbakhsh, G. Shoja, and V. King. Social-greedy: a sociallybased greedy routing algorithm for delay tolerant networks. In
MobiOpp, pages 159-162, 2010.
[13] R. Kleinberg. Geographic routing using hyperbolic space. In
INFOCOM, pages 1902-1909, 2007.
[14] G. Linden, B. Smith, and J. York. Industry report: Amazon.com
recommendations: Item-to-item collaborative filtering. IEEE
Distributed Systems Online, 4(1), 2003.
[15] S. M. McNee, J. Riedl, and J. A. Konstan. Being accurate is not
enough: how accuracy metrics have hurt recommender systems.
In CHI Extended Abstracts, pages 1097-1101, 2006.
[16] B. M. Sarwar, G. Karypis, J. A. Konstan, and J. Riedl. Item-based
collaborative filtering recommendation algorithms. In WWW,
pages 285-295, 2001.
[17] E. Spertus, M. Sahami, and O. Buyukkokten. Evaluating similarity measures: a large-scale study in the Orkut social network.
In KDD, pages 678-684, 2005.
[18] S. Spiekermann, M. Rothensee, and M. Klafft. Street marketing:
how proximity and context drive coupon redemption. Journal of
Consumer Marketing, 28(4):280-289, 2011.
[19] X. Su and T. M. Khoshgoftaar. A survey of collaborative filtering
techniques. Adv. Artificial Intellegence, 2009, 2009.
[20] P. K. Tat. A motivation-based model of coupon usage. Journal
of Promotion Management, 3(1-2): 1-14, 1996.
[21] J. Wang, A. P. de Vries, and M. J. T. Reinders. Unifying
user-based and item-based collaborative filtering approaches by
similarity fusion. In SIGIR, pages 501-508, 2006.
[22] M. Zhang and N. Hurley. Avoiding monotony: improving the
diversity of recommendation lists. In RecSys, pages 123-130,
2008.
232